Welcome, Guest. Please login or register.

Author Topic: 68040 particularities  (Read 2691 times)

Description:

0 Members and 1 Guest are viewing this topic.

Offline AmigaManceTopic starter

  • Hero Member
  • *****
  • Join Date: Apr 2005
  • Posts: 1278
    • Show only replies by AmigaMance
68040 particularities
« on: November 28, 2006, 12:21:07 PM »
Hello.
 I have notice a couple of things since i upgraded from a Blizzard 1230IV at 50mhz to a BlizzardPPC with a 68040 at 25mhz CPU:
 First, is that the 68040 appears to be MUUUUCH more cache-depended for its performance. For example, if you disable both caches on a 68030 it will become slower of course, but still retain a relatively decent performance. If you do the same on a 68040, it becomes extremely slow. it feels like a 68020/14mhz or even slower.
  Ok, i know that the caches of a 68040 are larger, but that doesn't explain it. Is there a technical reason for this?
 
 My second observation is that the speed benefit of moving from a 68030/50 to a 68040/25 is minimal. Sometimes hardly noticed. Perhaps because of the particularly limited memory bandwith of the BPPC? (IF this is not just a rumour)
A1200 PPC user.
 

Offline Piru

  • \' union select name,pwd--
  • Hero Member
  • *****
  • Join Date: Aug 2002
  • Posts: 6946
    • Show only replies by Piru
    • http://www.iki.fi/sintonen/
Re: 68040 particularities
« Reply #1 on: November 28, 2006, 12:24:33 PM »
Quote
Is there a technical reason for this?

Well, the most obvious reason would be that the 68030 runs at 2x clock speed.
 

Offline Vulture

  • Sr. Member
  • ****
  • Join Date: Nov 2005
  • Posts: 310
    • Show only replies by Vulture
Re: 68040 particularities
« Reply #2 on: November 28, 2006, 12:30:22 PM »
still a 68040@25 is supposedely quite faster than a 030@50, that's what benchmarks and rendering apps show in real life. I have no idea why the caches are so important to it tho....
 

Offline Piru

  • \' union select name,pwd--
  • Hero Member
  • *****
  • Join Date: Aug 2002
  • Posts: 6946
    • Show only replies by Piru
    • http://www.iki.fi/sintonen/
Re: 68040 particularities
« Reply #3 on: November 28, 2006, 12:36:30 PM »
Quote
still a 68040@25 is supposedely quite faster than a 030@50, that's what benchmarks and rendering apps show in real life.

...when caches are enabled.
 

Offline Vulture

  • Sr. Member
  • ****
  • Join Date: Nov 2005
  • Posts: 310
    • Show only replies by Vulture
Re: 68040 particularities
« Reply #4 on: November 29, 2006, 02:33:31 PM »
yes, that's understandable, but that was the original question, why is 040 such a slouch when the caches are off?
 

Offline Piru

  • \' union select name,pwd--
  • Hero Member
  • *****
  • Join Date: Aug 2002
  • Posts: 6946
    • Show only replies by Piru
    • http://www.iki.fi/sintonen/
Re: 68040 particularities
« Reply #5 on: November 29, 2006, 02:52:26 PM »
Well, first of all 68040 branch prediction (speculative execution) is turned off. It hurts performance quite a bit.

Second, all memory accesses (reads, writeback) will go to memory directly, without making use of cache. This will make the CPU stall a lot, waiting for memory accesses to finish.

Third, due to direct memory acceses it is likely that the 68040 pipeline (which is deeper than 030 one) can't be kept full all the time (that is the CPU is starved).

Fourth, 68040 improved performance depends quite a bit on copyback cache. Disable this and the performance is hurt a lot.

In short, 68040 depends on the cache to move the data cacheline at a time (16 bytes at a time) and to remain efficient. Remove this, and the CPU is limping badly. I doubt the 68040 was really optimized for the case where caches are deliberately turned off.

68030 on the other hand is basically 68020 + datacache + memory management unit. It's far simpler design and doesn't depend on the cache that much.
 

Offline adolescent

  • Hero Member
  • *****
  • Join Date: Sep 2003
  • Posts: 3056
    • Show only replies by adolescent
Re: 68040 particularities
« Reply #6 on: November 29, 2006, 03:10:28 PM »
Quote

AmigaMance wrote:

 My second observation is that the speed benefit of moving from a 68030/50 to a 68040/25 is minimal. Sometimes hardly noticed. Perhaps because of the particularly limited memory bandwith of the BPPC? (IF this is not just a rumour)


What tasks are you performing to observe this?  When I switched from 030/25MHz to 040/35MHz the speed difference was obvious in CPU intensive tasks like file packing/unpacking, image conversion, etc.  In straight WB use there wasn't much change (ie. it didn't speed up disk I/O much, and un-patched/un-optomised applications didn't seem to run much quicker.)
Time to move on.  Bye Amiga.org.  :(
 

Offline AmigaManceTopic starter

  • Hero Member
  • *****
  • Join Date: Apr 2005
  • Posts: 1278
    • Show only replies by AmigaMance
Re: 68040 particularities
« Reply #7 on: November 30, 2006, 01:11:04 PM »
Quote
Quote
AmigaMance wrote:

My second observation is that the speed benefit of moving from a 68030/50 to a 68040/25 is minimal. Sometimes hardly noticed. Perhaps because of the particularly limited memory bandwith of the BPPC? (IF this is not just a rumour)

 What tasks are you performing to observe this? When I switched from 030/25MHz to 040/35MHz the speed difference was obvious in CPU intensive tasks like file packing/unpacking, image conversion

 We can not compare dissimilar things. Your case is much different from mine.
A1200 PPC user.
 

Offline Dennis

  • Full Member
  • ***
  • Join Date: Dec 2005
  • Posts: 128
    • Show only replies by Dennis
Re: 68040 particularities
« Reply #8 on: November 30, 2006, 01:22:46 PM »
Quote
My second observation is that the speed benefit of moving from a 68030/50 to a 68040/25 is minimal. Sometimes hardly noticed. Perhaps because of the particularly limited memory bandwith of the BPPC? (IF this is not just a rumour)


I did exactly the same upgrade as you did when I still had my 1200. But to my experience, the 040@25 often felt even slower than the 030@50. Also, if i remember correctly, some 3D games like breathless or ab3d also ran slower on the 040. I still regret selling my blizzard1230-IV. It was the best accelerator I ever owned. After that one, it went downhill. The blizzardPPC/cybervision never worked well in my system (lots of ppc crashes, firmware loss etc..), and the first cyberstorm MKIII in my 3000 died after one year. The second ("repaired" by DCE) died after 3 hours....

Dennis
 

Offline srg86

  • Full Member
  • ***
  • Join Date: Aug 2004
  • Posts: 211
    • Show only replies by srg86
    • http://www.aopp12.dsl.pipex.com
Re: 68040 particularities
« Reply #9 on: November 30, 2006, 03:25:50 PM »
Most heavily pipelined processors are all depenend on their cache for performance as comparetively DRAM is very slow. The effect you're seeing on the 68040 is the same on the 68060, 486, Pentium, Athlon etc etc. Turn off your PC processor's L1 and L2 cahces and watch it crawl!
 

Offline alewis

  • Full Member
  • ***
  • Join Date: Jul 2006
  • Posts: 176
    • Show only replies by alewis
Re: 68040 particularities
« Reply #10 on: December 01, 2006, 09:57:55 AM »
My experience a bit different. Going from a B2000 with GVP 68030/40MHz to an A4000/040, the A4K felt nippier in day-to-day use. Going to a WarpEngine@40MHz blew the B2000 out of the water. of course, this was [probably] more due to the on-board DMA SCSI, faster (high data density, lower latency) SCSI drive, and then batter gfx card.

Going from a Warp040 to a CyberstormPPC/060, well, there is hardly any difference at all other than disk performance (which should be given its UW-SCSI3 vs SCSI-2)

Warp SCSI returns avg 8MB/sec
PPC SCSI retunrs avg 28MB/sec...

Same drive, limited by the 8bit narrow SCSI bus on the Warp.

But in day-to-day use, it actually feels slower. SysInfo's limited benchmark, iirc, shows it to be about 15% faster... TBH I don't consider it worth what I paid for it, even the "benefit" of a 16bit UW SCSI bus...
 

Offline vmc_de

  • Newbie
  • *
  • Join Date: Jun 2006
  • Posts: 4
    • Show only replies by vmc_de
    • http://www.vmc.de
Re: 68040 particularities
« Reply #11 on: December 01, 2006, 10:48:14 AM »
Hello Dennis,

please drop me a mail or read your pm box here
at amiga.org :)

Keep up the good work :)

Best regards,

Harald Frank
 

Offline Dennis

  • Full Member
  • ***
  • Join Date: Dec 2005
  • Posts: 128
    • Show only replies by Dennis
Re: 68040 particularities
« Reply #12 on: December 01, 2006, 11:15:34 AM »
Quote
My experience a bit different. Going from a B2000 with GVP 68030/40MHz to an A4000/040, the A4K felt nippier in day-to-day use. Going to a WarpEngine@40MHz blew the B2000 out of the water. of course, this was [probably] more due to the on-board DMA SCSI, faster (high data density, lower latency) SCSI drive, and then batter gfx card.


Maybe the blizzardppc was not a very fast design due the bus sharing with the PPC. Also, the blizzard1230 was a very fast design at the time. Maybe the step from a "fast" 030 to a "slow" 040 is not that great a step. The cyberstorm MKIII in my 3000 was VERY fast however. Together with an atlas-IV drive and a cybervision64/3D it was a very nice setup. Too bad MKIII's don't live that long.  :-(

Dennis
 

Offline alewis

  • Full Member
  • ***
  • Join Date: Jul 2006
  • Posts: 176
    • Show only replies by alewis
Re: 68040 particularities
« Reply #13 on: December 01, 2006, 12:09:27 PM »
Hmm, well I have a Blizzard 060 in an A1200, and the experience is the same with that. Its fast, but not the "kick in the pants" step up from an 040. BTW, it's a CyberstormPPC/060 in the A4K, also has a Cybervision64/3D, and a 10K RPM U160 SCSI disk - which is a tad wasteful, but hey.

Quote

Dennis wrote:
Maybe the blizzardppc was not a very fast design due the bus sharing with the PPC. Also, the blizzard1230 was a very fast design at the time. Maybe the step from a "fast" 030 to a "slow" 040 is not that great a step. The cyberstorm MKIII in my 3000 was VERY fast however. Together with an atlas-IV drive and a cybervision64/3D it was a very nice setup. Too bad MKIII's don't live that long.  :-(

Dennis